Enhancing Forest Fire Severity Prediction Using Machine Learning: Evaluation of Linear Regression, Random Forest, and Gradient Boosting Models

Kumar, Shiyamal (2025) Enhancing Forest Fire Severity Prediction Using Machine Learning: Evaluation of Linear Regression, Random Forest, and Gradient Boosting Models. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (609kB) | Preview
Preview	PDF (Configuration Manual) Download (793kB) | Preview

Abstract

The threats of forest fires are severe and include ecological systems, animal and plant life, human health and property. Due to rising cases associated with the climate, land-use patterns, and environmental mismanagement, the necessity of high-quality forest fire severity predictions has reached its apex. This paper will investigate how machine learning regression algorithms or Linear Regression; Random Forest and Gradient Boosting can be employed to predict the intensity of forest fires based on some historical data regarding meteorology and environmental data. The study entailed an extensive data cleaning step, such as the treatment of missing values, normalization, feature engineering, and skewness transformation of variables. The major features like temperature, humidity, wind speed, and rainfall were studied and used to train the models. Performance of the models was measured in terms of R², MAE and RMSE values. Although the above preprocessing procedure was extremely rigorous and the models were trained with vast amounts of data, it was found that the models did not give very accurate predictions. Linear Regression posted the highest R² value of 0.0311 but none were significantly high, whereas the ensemble methods performed worse. The underestimation of the high severity fire events was consistently reflected through lack of capturing the high-order, non-linear nature of interactions of the fire behavior. The results indicate that the traditional regression models do not perform well in this particular area and there is a possibility to achieve better outcomes using more powerful approaches like deep learning or a combination of both. The paper presents the basis of future research on the development of efficient, real-time tools of forest fire risk management.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Horn, Christian UNSPECIFIED
Uncontrolled Keywords:	Forest Fire Prediction; Regression Models; Meteorological Data; Model Performance; Deep Learning
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences S Agriculture > SD Forestry G Geography. Anthropology. Recreation > GE Environmental Sciences > Environmental protection > Climate change mitigation Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning G Geography. Anthropology. Recreation > GE Environmental Sciences > Earth sciences > Atmospheric science > Meteorology
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	24 Aug 2026 12:20
Last Modified:	24 Aug 2026 12:20
URI:	https://norma.ncirl.ie/id/eprint/9605

Actions (login required)

View Item